fix(mcq): load JSON prediction tables for evaluation

[rudycelekli]

With PRED_FORMAT=json, native prediction exports contain record lists. MCQ evaluators read these through ordinary load(), then call DataFrame methods; scored prediction tables saved as JSON fail again when reloaded. Use the existing prediction-table reader for the tabular reads in text/image MCQ evaluation and their image overrides. Judge-cache dictionaries and ordinary JSON loading retain their existing semantics.

This complements #1172: its format/filename changes leave these MCQ reader calls unchanged. Merged #1729 wires the table-aware reader into inference resume; this change covers evaluation consumers. One grouped change covers31 source-audited table read sites.

Validation: native local TSV dataset constructors, actual prediction exports, original exact matcher/scorer producing native resume caches, and full cached evaluation/reporting. Six JSON input/reload cases fail against the base while eight controls pass. All14 focused cases pass after the fix; existing JSON-resume/file/NumPy neighbors bring the total to67 passes. XLSX, TSV, physical Pillow image prompts, generic text/image MCQ, and a local CVBench subtype are covered. Ordinary JSON dict/list return types remain unchanged. All nine pinned pre-commit hooks pass on materialized source/metadata.

Tests isolate unrelated optional model imports and execute exact shipping definitions. No model/provider calls, GPU benchmarks, downloaded datasets, benchmark accuracy, hosted CI, or validation of every specialized scoring branch is claimed. AI assistance: prepared with Codex and verified using native local fixtures.

[ssiq]

@rudycelekli Thank you for your contribution. load_prediction_file should only be used to load inference result files. It should not be used for files generated during evaluation, as those files are not guaranteed to contain a DataFrame or a prediction column.

[rudycelekli]

Thanks, agreed. Pushed 97bc271: load_prediction_file is now used only for reading the inference output (eval_file). The tables that evaluation itself writes and reloads (the _result / _vanilla_*_result / eval-record files) go through a new load_table_file in vlmeval/smp/file.py, which does not require a prediction column and only turns the record list that JSON serialization produces back into a DataFrame; any other loaded value is returned unchanged. The test now reloads the scored table with it. Local run: tests/test_mcq_prediction_tables.py and tests/test_json_resume.py pass (18 passed); full CI not yet re-checked.